>>106488987
Try prompting the age or features (chubby cheeks, large nose, etc) or even negging out facial features or ethnic features. Negging ethnicities can help a lot because of the heavy Korean influence in most base datasets.

>>106489821
Sometimes the model can latch on to something innocuous during training (I had a problem with lanyards many LoRAs ago), BFL probably didn't catch it because they were focused on other things (coherency, etc)